# @dreb/agent-core

Stateful agent with tool execution and event streaming. Built on `@dreb/ai`.

## Installation

```bash
npm install @dreb/agent-core
```

## Quick Start

```typescript
import { Agent } from "@dreb/agent-core";
import { getModel } from "@dreb/ai";

const agent = new Agent({
  initialState: {
    systemPrompt: "You are a helpful assistant.",
    model: getModel("anthropic", "claude-sonnet-4-20250514"),
  },
});

agent.subscribe((event) => {
  if (event.type === "message_update" && event.assistantMessageEvent.type === "text_delta") {
    // Stream just the new text chunk
    process.stdout.write(event.assistantMessageEvent.delta);
  }
});

await agent.prompt("Hello!");
```

## Core Concepts

### AgentMessage vs LLM Message

The agent works with `AgentMessage`, a flexible type that can include:
- Standard LLM messages (`user`, `assistant`, `toolResult`)
- Custom app-specific message types via declaration merging

LLMs only understand `user`, `assistant`, and `toolResult`. The `convertToLlm` function bridges this gap by filtering and transforming messages before each LLM call.

### Message Flow

```
settled AgentContext → beforeLlmCall() → transformContext() → convertToLlm() → LLM
                         (optional)          (optional)         (required)
```

1. **beforeLlmCall**: Inspect settled user/tool-result context and atomically replace messages and/or the active model before the request
2. **transformContext**: Prune old messages or inject external context for the request
3. **convertToLlm**: Filter out UI-only messages and convert custom types to LLM format

`beforeLlmCall` runs after `shouldContinue` permits a request and before `turn_start`. Returning `{ messages, model }` updates both the active loop and `Agent.state`; returning `undefined` preserves them. The hook receives a snapshot, so mutations without a returned replacement are ignored. With `Agent`, a thrown error terminates the run with a visible assistant error instead of applying a partial replacement; low-level loop callers must surface failures themselves and return `undefined` rather than reject.

## Event Flow

The agent emits events for UI updates. Understanding the event sequence helps build responsive interfaces.

### prompt() Event Sequence

When you call `prompt("Hello")`:

```
prompt("Hello")
├─ agent_start
├─ turn_start
├─ message_start   { message: userMessage }      // Your prompt
├─ message_end     { message: userMessage }
├─ message_start   { message: assistantMessage } // LLM starts responding
├─ message_update  { message: partial... }       // Streaming chunks
├─ message_update  { message: partial... }
├─ message_end     { message: assistantMessage } // Complete response
├─ turn_end        { message, toolResults: [] }
└─ agent_end       { messages: [...] }
```

### With Tool Calls

If the assistant calls tools, the loop continues:

```
prompt("Read config.json")
├─ agent_start
├─ turn_start
├─ message_start/end  { userMessage }
├─ message_start      { assistantMessage with toolCall }
├─ message_update...
├─ message_end        { assistantMessage }
├─ tool_execution_start  { toolCallId, toolName, args }
├─ tool_execution_update { partialResult }           // If tool streams
├─ tool_execution_end    { toolCallId, result }
├─ message_start/end  { toolResultMessage }
├─ turn_end           { message, toolResults: [toolResult] }
│
├─ turn_start                                        // Next turn
├─ message_start      { assistantMessage }           // LLM responds to tool result
├─ message_update...
├─ message_end
├─ turn_end
└─ agent_end
```

Tool execution mode is configurable:

- `parallel` (default): preflight tool calls sequentially, execute allowed tools concurrently, emit final `tool_execution_end` and `toolResult` messages in assistant source order
- `sequential`: execute tool calls one by one, matching the historical behavior

The `beforeToolCall` hook runs after `tool_execution_start` and validated argument parsing. It can block execution. The `afterToolCall` hook runs after tool execution finishes and before `tool_execution_end` and final tool result message events are emitted.

When you use the `Agent` class, assistant `message_end` processing is treated as a barrier before tool preflight begins. That means `beforeToolCall` sees agent state that already includes the assistant message that requested the tool call.

### continue() Event Sequence

`continue()` resumes from existing context without adding a new message. Use it for retries after errors.

```typescript
// After an error, retry from current state
await agent.continue();
```

The last message in context must be `user` or `toolResult` (not `assistant`).

### Event Types

| Event | Description |
|-------|-------------|
| `agent_start` | Agent begins processing; includes resolved model and effective thinking level |
| `agent_end` | Agent completes with all new messages |
| `turn_start` | New turn begins (one LLM call + tool executions) |
| `turn_end` | Turn completes with assistant message and tool results |
| `message_start` | Any message begins (user, assistant, toolResult) |
| `message_update` | **Assistant only.** Includes `assistantMessageEvent` with delta |
| `message_end` | Message completes |
| `stream_retry` | Provider stream dropped; retrying after discarding partial response |
| `length_retry` | Response truncated at the token limit; retrying at the configured output limit |
| `tool_execution_start` | Tool begins |
| `tool_execution_update` | Tool streams progress |
| `tool_execution_end` | Tool completes |

`agent_start` carries the provider/model identity and the effective thinking level sent to the provider for that run:

```typescript
{ type: "agent_start", model: { provider: "anthropic", id: "claude-sonnet-4-6" }, thinkingLevel: "high" }
```

## Agent Options

```typescript
const agent = new Agent({
  // Initial state
  initialState: {
    systemPrompt: string,
    model: Model<any>,
    thinkingLevel: "off" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max",
    tools: AgentTool<any>[],
    messages: AgentMessage[],
  },

  // Prepare settled context immediately before each LLM request. Both fields
  // are optional; replacements apply to the active loop and Agent.state.
  beforeLlmCall: async (context, signal) => ({
    messages: compactedMessages,
    model: resizedContextModel,
  }),

  // Convert AgentMessage[] to LLM Message[] (required for custom message types)
  convertToLlm: (messages) => messages.filter(...),

  // Transform context before convertToLlm (for pruning, compaction)
  transformContext: async (messages, signal) => pruneOldMessages(messages),

  // Steering mode: "one-at-a-time" (default) or "all"
  steeringMode: "one-at-a-time",

  // Follow-up mode: "one-at-a-time" (default) or "all"
  followUpMode: "one-at-a-time",

  // Custom stream function (for proxy backends)
  streamFn: streamProxy,

  // Session ID for provider caching
  sessionId: "session-123",

  // Dynamic API key resolution (for expiring OAuth tokens)
  getApiKey: async (provider) => refreshToken(),

  // Tool execution mode: "parallel" (default) or "sequential"
  toolExecution: "parallel",

  // Preflight each tool call after args are validated. Can block execution.
  beforeToolCall: async ({ toolCall, args, context }) => {
    if (toolCall.name === "bash") {
      return { block: true, reason: "bash is disabled" };
    }
  },

  // Postprocess each tool result before final tool events are emitted.
  afterToolCall: async ({ toolCall, result, isError, context }) => {
    if (!isError) {
      return { details: { ...result.details, audited: true } };
    }
  },

  // Forward non-fatal provider warnings (e.g. malformed SSE/JSON) to the caller
  onWarning: (code, message) => console.warn(`[${code}] ${message}`),

  // Stream drop retry: number of retries when a provider stream ends without
  // its terminal event (default: 3). Set to 0 to disable.
  streamRetries: 3,

  // Base delay for exponential backoff between stream retries in ms (default: 1000)
  streamRetryBaseDelayMs: 1000,

  // Length retry: number of retries when a turn ends with stopReason "length"
  // (model exhausted its configured output limit). Each retry uses that same limit.
  // Separate from streamRetries (default: 2). Set to 0 to disable.
  lengthRetries: 2,

  // Custom thinking budgets for token-based providers
  thinkingBudgets: {
    minimal: 128,
    low: 512,
    medium: 1024,
    high: 2048,
  },

  // Thinking display for adaptive-thinking Anthropic/Bedrock Claude models.
  // "summarized" returns thinking text; "omitted" returns only an encrypted
  // signature (lower latency). Ignored by other models. Opus 4.7+ default to
  // "omitted" at the API, so pass "summarized" to keep thinking visible.
  thinkingDisplay: "summarized",
});
```

## Agent State

```typescript
interface AgentState {
  systemPrompt: string;
  model: Model<any>;
  thinkingLevel: ThinkingLevel;
  tools: AgentTool<any>[];
  messages: AgentMessage[];
  isStreaming: boolean;
  streamMessage: AgentMessage | null;  // Current partial during streaming
  pendingToolCalls: Set<string>;
  error?: string;
}
```

Access via `agent.state`. During streaming, `streamMessage` contains the partial assistant message.

## Methods

### Prompting

```typescript
// Text prompt
await agent.prompt("Hello");

// With images
await agent.prompt("What's in this image?", [
  { type: "image", data: base64Data, mimeType: "image/jpeg" }
]);

// AgentMessage directly
await agent.prompt({ role: "user", content: "Hello", timestamp: Date.now() });

// Continue from current context (last message must be user or toolResult)
await agent.continue();
```

### State Management

```typescript
agent.setSystemPrompt("New prompt");
agent.setModel(getModel("openai", "gpt-4o"));
agent.setThinkingLevel("medium");
agent.setTools([myTool]);
agent.setToolExecution("sequential");
agent.setBeforeLlmCall(async (context, signal) => ({ messages: context.messages }));
agent.setBeforeToolCall(async ({ toolCall }) => undefined);
agent.setAfterToolCall(async ({ toolCall, result }) => undefined);
agent.replaceMessages(newMessages);
agent.appendMessage(message);
agent.clearMessages();
agent.reset();  // Clear everything
```

### Session and Thinking Budgets

```typescript
agent.sessionId = "session-123";

agent.thinkingBudgets = {
  minimal: 128,
  low: 512,
  medium: 1024,
  high: 2048,
};
```

### Control

```typescript
agent.abort();           // Cancel current operation
await agent.waitForIdle(); // Wait for completion
```

### Events

```typescript
const unsubscribe = agent.subscribe((event) => {
  console.log(event.type);
});
unsubscribe();
```

## Steering and Follow-up

Steering messages let you interrupt the agent while tools are running. Follow-up messages let you queue work after the agent would otherwise stop.

```typescript
agent.setSteeringMode("one-at-a-time");
agent.setFollowUpMode("one-at-a-time");

// While agent is running tools
agent.steer({
  role: "user",
  content: "Stop! Do this instead.",
  timestamp: Date.now(),
});

// After the agent finishes its current work
agent.followUp({
  role: "user",
  content: "Also summarize the result.",
  timestamp: Date.now(),
});

const steeringMode = agent.getSteeringMode();
const followUpMode = agent.getFollowUpMode();

agent.clearSteeringQueue();
agent.clearFollowUpQueue();
agent.clearAllQueues();
```

Use clearSteeringQueue, clearFollowUpQueue, or clearAllQueues to drop queued messages.

When steering messages are detected after a turn completes:
1. All tool calls from the current assistant message have already finished
2. Steering messages are injected
3. The LLM responds on the next turn

Follow-up messages are checked only when there are no more tool calls and no steering messages. If any are queued, they are injected and another turn runs.

## Custom Message Types

Extend `AgentMessage` via declaration merging:

```typescript
declare module "@dreb/agent-core" {
  interface CustomAgentMessages {
    notification: { role: "notification"; text: string; timestamp: number };
  }
}

// Now valid
const msg: AgentMessage = { role: "notification", text: "Info", timestamp: Date.now() };
```

Handle custom types in `convertToLlm`:

```typescript
const agent = new Agent({
  convertToLlm: (messages) => messages.flatMap(m => {
    if (m.role === "notification") return []; // Filter out
    return [m];
  }),
});
```

## Tools

Define tools using `AgentTool`:

```typescript
import { Type } from "@sinclair/typebox";

const readFileTool: AgentTool = {
  name: "read_file",
  label: "Read File",  // For UI display
  description: "Read a file's contents",
  parameters: Type.Object({
    path: Type.String({ description: "File path" }),
  }),
  execute: async (toolCallId, params, signal, onUpdate) => {
    const content = await fs.readFile(params.path, "utf-8");

    // Optional: stream progress
    onUpdate?.({ content: [{ type: "text", text: "Reading..." }], details: {} });

    return {
      content: [{ type: "text", text: content }],
      details: { path: params.path, size: content.length },
    };
  },
};

agent.setTools([readFileTool]);
```

### Error Handling

**Throw an error** when a tool fails. Do not return error messages as content.

```typescript
execute: async (toolCallId, params, signal, onUpdate) => {
  if (!fs.existsSync(params.path)) {
    throw new Error(`File not found: ${params.path}`);
  }
  // Return content only on success
  return { content: [{ type: "text", text: "..." }] };
}
```

Thrown errors are caught by the agent and reported to the LLM as tool errors with `isError: true`.

## Proxy Usage

For browser apps that proxy through a backend:

```typescript
import { Agent, streamProxy } from "@dreb/agent-core";

const agent = new Agent({
  streamFn: (model, context, options) =>
    streamProxy(model, context, {
      ...options,
      authToken: "...",
      proxyUrl: "https://your-server.com",
    }),
});
```

## Low-Level API

For direct control without the Agent class:

```typescript
import { agentLoop, agentLoopContinue } from "@dreb/agent-core";

const context: AgentContext = {
  systemPrompt: "You are helpful.",
  messages: [],
  tools: [],
};

const config: AgentLoopConfig = {
  model: getModel("openai", "gpt-4o"),
  beforeLlmCall: async (settledContext, signal) => ({ messages: settledContext.messages }),
  convertToLlm: (msgs) => msgs.filter(m => ["user", "assistant", "toolResult"].includes(m.role)),
  toolExecution: "parallel",
  beforeToolCall: async ({ toolCall, args, context }) => undefined,
  afterToolCall: async ({ toolCall, result, isError, context }) => undefined,
};

const userMessage = { role: "user", content: "Hello", timestamp: Date.now() };

for await (const event of agentLoop([userMessage], context, config)) {
  console.log(event.type);
}

// Continue from existing context
for await (const event of agentLoopContinue(context, config)) {
  console.log(event.type);
}
```

These low-level streams are observational. They preserve event order, but they do not wait for your async event handling to settle before later producer phases continue. `beforeLlmCall` itself is awaited and its returned replacement is applied atomically before the request. If you need ordinary message-event processing to act as a barrier before tool preflight, use the `Agent` class instead of raw `agentLoop()` or `agentLoopContinue()`.

## License

MIT